A novel pipelined end-to-end relation extraction framework with entity mentions and contextual semantic representation

The mainstream method of end-to-end relation extraction is to jointly extract entities and relations by sharing span representation, which, however, may cause feature conflict. The advent of advanced pre-trained models enhances the ability to learn span semantic representation and allows the breaking of the dominance of joint models. We argue the benefits of using separate encoders for entity recognition and relation classification and propose a novel pipelined end-to-end relation extraction framework. By adopting attention mechanisms, the framework has the ability to fuse contextual semantic representation, which is missed in other pipelined models. By introducing explicit entity mentions, the framework is able to capture entities' location information and type information, which are difficult to utilize in joint models. Several elaborate tricks are integrated into the training process of the framework to further improve its performance. Our experiments show that our method increases the state-of-the-art relation F1-score on CoNLL04, ADE and SciERC datasets to 75.6% (+1.2%), 85.0% (+1.2%), 43.9% (+2.3%), respectively, indicating that our pipelined approach is promising in end-to-end relation extraction.